Are Protective Labor Market Institutions Really at the Root of Unemployment? A Critical Perspective on the Statistical Evidence 1

Transcription

1 July 14, 2006 Are Protective Labor Market Institutions Really at the Root of Unemployment? A Critical Perspective on the Statistical Evidence 1 David R. Howell, Dean Baker, Andrew Glyn and John Schmitt It is widely accepted that the rigidities created by labor market institutions explain the pattern of unemployment across countries. A rapidly expanding recent literature has explored the statistical support for this orthodox view. This paper offers a critical perspective on the evidence. We focus on the protective institutions that are the usual suspects: unemployment benefit entitlements, employment protection laws, and trade unions. Given the dominance of this view, the simple correlation evidence offers remarkably little support. The most robust finding of the cross-country regression literature points to a potentially important role for unemployment benefits generosity, but there are reasons to doubt the strength of this relationship and even the direction of causation. The micro evidence on the effects of major changes in benefit generosity on the duration of unemployment (and the exit rate into employment) is much less supportive of a sizable impact of benefit generosity on the aggregate unemployment rate than is often suggested. Finally, we find little evidence to suggest that changes in the strength of these protective labor market institutions can explain either the success of the success stories or the continued high unemployment of the four large continental European countries.

2 1. Unemployment and Institutions: The Basic Facts 2. Measurement Issues 2.1 Unemployment 2.2 Labor Market Institutions 3. Correlation Evidence 3.1 Casual Associations 3.2 Simple Correlations 4. Macroeconometric Evidence 4.1 The Consensus View: Centrality of Labor Market Institutions 4.2 Early OECD Studies 4.3 Other Early Cross-Country Studies 4.4 Explaining Changes over Time with Annual Data 4.5 Time Series Evidence from the UK 4.6 Assessment 5. Unemployment Benefit Compensation and Unemployment 5.1 Some Initial Considerations 5.2 Interpretation of Benefit Effects 5.3 The Microeconometric Evidence Evidence from Literature Surveys Micro Evidence on Unemployment Duration 5.4 Assessment 6. Comprehensive Labor Market Reform and Unemployment 6.1 Nickell s Reforms Index 6.2 OECD Reforms Indices 7. Conclusion 2

3 3 As recently as 1979, only Ireland and Portugal among the 20 most developed (OECDmember) countries reported unemployment rates above 8 percent (each at about 8.5%). Just four years later, 11 of these 20 countries posted higher rates and six reached double-digit levels, ranging from Belgium (10.7%) to Ireland (14.9%). This collapse in employment performance persisted throughout the 1980s and 1990s. Between 1995 and 1997, as the U.S. was showing rates between 5.6 to 4.9 percent, OECD-Europe ranged from 10.1 to 9.7 percent. By 2005, the OECD-Europe rate had dropped to 8.6 percent, but both core economies of continental Europe, France and Germany, had rates of 9.5 percent. 2 Much like the response of economists to the Great Depression, the dominant explanation for persistent high unemployment has centered on supply-side rigidities generated by protective labor market institutions, and correspondingly, the proposed solution has been greater (downward) wage flexibility and stronger work incentives. As Fitoussi (2006) has put it, The reference model, in the plea for structural reforms, is centered on an economy with perfect competition and rational expectations. In such a model full employment is always assured absent rigidities.... Spurred in particular by the influence of the Layard, Nickell and Jackman (1991) and the OECD s Jobs Study (1994), this orthodox rigidity account thoroughly ruled the field by the late 1990s with quite straightforward policy implications. The IMF (2003, p. 129) points out that leading international institutions the IMF, OECD and the European Commission have long argued that the causes of unemployment can be found in labor market institutions. Accordingly, countries with high unemployment have been repeatedly urged to undertake comprehensive structural reforms to reduce labor market rigidities. This view has become so widely accepted that a leading scholar could recently claim in the Journal of Economic Perspectives that evidence supports the traditional view that rigidities that reduce competition in labor markets are typically responsible for high unemployment without actually citing any peer-reviewed evidence (St. Paul, 2004, p. 53). Three labor market institutions have been held to play the premier roles in the promotion of employment-unfriendly rigidities: unemployment benefit entitlements, employment protection laws, and trade unions. 3 Not coincidentally, these are the key institutional mechanisms most developed countries have relied upon to shelter less-skilled workers from the most harmful effects of competitive labor markets. We will refer to them as

4 4 protective labor market institutions (PLMIs) and distinguish them from other key institutions that have important labor market effects. These include active labor market policies (ALMP), which are concerned with matching and preparing workers for jobs; tax policy, which influences behavior and affects labor costs, but is principally designed to raise revenue, not protect workers; and housing policies, which affect ownership rates and could affect worker mobility, but are not designed to protect workers as workers. The dominance of the orthodox rigidities explanation of unemployment and the recent focus on macroeconometric testing reflects a striking evolution in mainstream economics. As recently as 1994, Charles Bean s influential survey of European unemployment allocated little space to evidence on the effects of these key PLMIs on employment performance, finding little compelling empirical support in the literature for any of them. 4 Bean concluded with three recommendations for future research, the first of which was to discourage macroeconometric testing: There is simply not enough information in the data to give clear signals on the relative merits of the competing hypotheses (p. 615). Ignoring this advice, cross-country macroeconometric studies have expanded at an increasing rate (e.g., see OECD, 2006; Blanchard, 2006). This paper critically assesses the empirical evidence on the effects of labor market institutions on the cross country pattern of unemployment, focusing on the role played by the three core protective institutions unemployment benefits, employment protection, and trade unions. After outlining the basic facts on the cross-country pattern of unemployment and labor market institutions, section 2 considers some issues of measurement. Section 3 then evaluates the simple correlation evidence between standard measures of labor market institutions and unemployment. Section 4 addresses the macroeconometric evidence. Since the most robust evidence in favor of the orthodox rigidity view concerns the role played by unemployment benefit generosity, Section 5 takes a closer look at the interpretation of benefit effects in the macroeconometric research. It also reviews the microeconometric evidence, which has often been cited as supporting evidence. Section 6 then assesses recent efforts to develop aggregate indicators of labor market reform with the goal of showing the payoff of comprehensive labor market reform for employment performance. We conclude in Section 7 with a summary and a brief discussion of the interplay between theory, evidence and policy recommendations.

5 5 1. Unemployment and Institutions: The Basic Facts Figure 1 shows the levels and dispersion of unemployment rates for 19 OECD-member countries for each 5-year period between 1960 and 2004, and includes the most recent figures for 2005 at the far right. As a reference, the line that runs from left to right marks the U.S. rate. The table at the bottom presents the U.S. rate, the median, and a measure of the dispersion of rates (the standard deviation). This figure highlights some key facts about the changing nature of the unemployment problem in the developed world. First, nearly all countries experienced escalating unemployment through at least the late 1980s. The median unemployment rate (see the table below the Figure) rose from 1.9% in the late 1960s to 8.8% in Second, the dispersion of rates has moved upward with the median. The standard deviation for these 19 countries increased sharply from the range in the 1960s-70s to in the 1980s-90s. Third, unemployment rates have declined and converged substantially since the late 1990s: the median fell from 7.9% in to 5.3% in and 5.2% for 2005; the standard deviation fell from 3.9 to just below 2, which is about where it was on average in the 1970s. The figure shows that the distribution of unemployment rates in 2005 falls in a range of about six percentage points (from four to ten percent), about the same as the range in (from about zero to 6 percent). And fourth, the unemployment performance of the U.S. varies dramatically over this period, from among the countries with the very highest rates through the first two decades ( ) to among those with the lowest rates in the second half of the 1990s, and back again to close to the median since 2000 ( and 2005). It is also worth noting that New Zealand has regained its position as the country with the lowest unemployment rate; Ireland has dropped to the second lowest rate from the second highest in ; and Spain as experienced a remarkable decline, to a level that is now just below that of Germany and France. In the popular press and in a surprising number of professional papers, Europe is often portrayed as a single entity characterized by high unemployment and strong social protections, in contrast to the much better performing and relatively unregulated labor markets of the U.S. and other Anglo-Saxon economies. This conventional view greatly misrepresents the facts. Table 1 provides unemployment rates for 2003 by demographic

6 6 group for countries in three groupings: six English-speaking countries with generally low unemployment (Canada remains at higher levels); six high unemployment continental European countries; and six European low unemployment countries. This table shows that the six liberal, English speaking countries had average unemployment rates nearly identical to those of the six low-unemployment European countries for all four demographic groups male and female young and prime age workers. The five high-unemployment Continental countries show substantially higher unemployment for each age-gender group. With the exception of Germany, each has experienced extremely high youth unemployment. Female youth show rates of 17.5% in Belgium, 22.8% in France, 27.2% in Spain, and 30.9% in Italy; male youth rates range from 18-23%. Clearly, young people in these four countries account for an important part of the European unemployment problem. It should be recognized, however, that using an alternative measure of unemployment as a share of the youth population rather than as a share of the youth labor force the picture looks quite a bit different. With this alternative measure, for example, France and the U.S. have similar youth unemployment rates (Howell, 2005, chapter 1). The similarity between unemployment rates for the liberal English-speaking countries and low-unemployment Europe is notable because the latter remain characterized by strong welfare states and highly protective labor market institutions. Table 2 shows that while both of the European groups are characterized by much higher levels of social protection and regulation (rows 4-9) and much higher tax revenue shares (row 10), only the conservative/corporatist economies of high-unemployment Europe show worse employment performance than the liberal economies. Indeed, on both unemployment and employment rates, the northern European welfare states show, on average, superior labor market performance to the liberal ones (rows 1-3), and they do so with much lower wage inequality (row 11). As Nickell (1997; 2003) has pointed out, many Europeans live in regions with lower unemployment rates than the U.S. and most of the unemployed of Europe live in four large countries (France, Spain, Italy and Germany). 2. Measurement Issues

7 7 Before examining considering the statistical evidence on the relationships between unemployment and labor market institutions, it is worth taking a brief look at the construction and quality of the measures. While the literature has been characterized by a steady increase in the sophistication of econometric techniques, remarkably little attention has been paid to the quality and consistency of the data. This is particularly curious, since, thanks largely to the efforts of the OECD, there have been impressive improvements in both the quality of the institutional measures and the consistency of the unemployment series. 2.1 Unemployment For the most part, empirical work on unemployment has proceeded under the assumption that the dependent variable the unemployment rate is well-measured and comparable ( harmonized or standardized ) over time and across countries. Indeed, rarely do authors offer more than simply a citation for the source of the variable. But a closer look at the data shows that comparability is often limited and that the many different series in use can produce quite different results. The source data are collected at the national level, both as registered unemployment (collected by the national employment service) and from household surveys (similar to the U.S. Current Population Survey). Over time, OECD countries have adopted international standards that establish the criteria for who is unemployed based on household surveys, but some series refer to those between (for the U.S. it is age 16) and others to those over age 15. While comparability has increased as data collection and processing methods and criteria have converged across countries, both the OECD and the U.S. Bureau of Labor Statistics have developed additional series that attempt to make the rates more comparable (Sorrentino, 2000). But full comparability remains elusive. Since the unemployed cannot be employed, how a respondent replies to the question asking whether or not he/she was employed for at least an hour for pay in the reference week will reflect to some degree local social norms and levels of economic development (Howell, 2005). What is considered real employment may differ substantially across regions and countries, and this may help explain, for example, how Mexico and the U.S. could have similar unemployment rates, calculated with similar definitions and methods (Howell, 2005; Martin, 2000).

8 8 In the early unemployment-institutions tests by Layard, Nickell and Jackman (1991) a simple cross section of unemployment rates averaged over the period was used. As the literature has progressed, analysts have demanded annual series covering a much longer period, often stretching back to the early 1960s, and aimed at explaining changes in unemployment over time with changes in institutional and policy measures, with fixed country effects (see section 4). This empirical strategy relies heavily on the use of historically consistent unemployment rate series for each country. 5 For many countries, this consistency criterion is clearly not satisfied. National methods have changed substantially over this period and there is no standardized or even historically consistent series that comes close to dating back even to the 1970s for many of the 19 or so OECD countries that usually appear in the cross-country tests. For example, the OECD s standardized data extend back to 1980 for just 9 OECD countries. The OECD s longest historical unemployment series (not standardized) is available for just 9 countries for 1970 and 4 countries for Our attempt to determine the change in the OECD s unemployment rate for The Netherlands over the last two decades offers an illustration of the problem with the historical statistics. The Bassanini-Duval (2006) macroeconometric tests provide the empirical basis for Chapter 7 of the new OECD Employment Outlook (2006). This research focuses on the period, and in both the Bassanini-Duval report and the Employment Outlook Chapter, a cross-country scatter plot of predicted against observed changes in unemployment is presented for (OECD 2006, Figure 7.3). The conclusion drawn from this correlation is that a small number of labor market policies and institutions can largely account for cross-country differences in how unemployment has evolved since the early 1980s. The Bassanini-Duval figure shows a decline in unemployment over the period for Holland of over 9 percentage points, based on a 1982 rate of 13.2 percent. The referenced source is the OECD s Labour Force Statistics, but in the OECD s Labour Force Statistics publication (OECD, 2004) the change is just 7.4 percentage points (11.6% in 1982 to 4.2% in 2003). Another OECD series the standardized rates that appear in the OECD s Employment Outlook - shows a still smaller change of 4 points (7.68 to 3.68). By way of comparison, the series used by the IMF (2003) and Nickell et al. (2001) shows a 4.8 point change (8.5 to 3.7). 7 The poorly performing countries are located at the other end of the

9 9 Bassanini-Duval figure. The worst among these is Switzerland, which shows an increase in observed unemployment for of about 4 percentage points. The problem is that it is computed on the basis of an implausibly low 1982 unemployment rate of 0.2 percent. Prior to 1991, the Swiss data referred only to registered unemployment, and for this reason the OECD publishes no standardized figures for Switzerland prior to that date. In short, the 4 point increase in Swiss unemployment reflects a comparison of a registered rate of 0.2 percent in 1982 and a household survey based rate of 4.2 percent in The change between them has little meaning, either for changes in employment performance within the country or for cross-country comparisons. Switzerland is by no means unique in relying on administrative data for historical time series. 8 The Dutch and Swiss are extreme examples that illustrate the weaknesses in the historical time series on unemployment for many countries. The comparability of the unemployment rate numbers declines substantially the further back the time series runs it is not until the early 1990s that nearly all major OECD member countries generally adopted the ILO standard (the broad definition of unemployment based on household surveys). But even for recent years, differences remain both across countries and over time within countries over exactly how the ILO unemployment rate is calculated (such as what qualifies as active job search), which may have important effects on the calculated rate. At a minimum, since the quality and consistency of the data have evolved and the use of different series can produce different results, it seems appropriate for studies to provide more documentation and justification for their data. 2.2 Labor Market Institutions Statistical tests of the effects of labor market institutions on the pattern of unemployment required the development of measures of institutions and policies. This effort was pioneered by Nickell and Layard, whose measures appeared in a series of papers and books in the earlyand mid-1990s (Layard, Nickell, and Jackman, 1991 and 1994; Nickell and Bell, 1994; Layard and Nickell, 1996). Considerable subjective judgment was required for many of these inherently difficult-to-measure institutions. For example, the measure of unemployment benefits duration that was employed extensively in the 1990s was an estimate of the number of years a representative unemployed

10 10 worker was eligible for benefits. Thus, Layard et al. (1994, p. 74) gave the U.S. a score of.5, Denmark 2.5, and France With eight other countries, The Netherlands received a 4, indicating indefinite duration of benefits. In their survey of the benefit entitlement literature, Atkinson and Micklewright (1991) single out these data for criticism, pointing out that the institutional design of each of the countries with indefinite duration scores are quite different, and these differences have substantial effects on how generous the systems really are for which parts of the unemployed population. 9 To take one example, Atkinson and Micklewright (Table 3, p. 1696) explain that the reality behind the indefinite score for the Netherlands in the mid-1980s is considerably more complicated and certainly not indefinite: UI at 70% of last earned wage for between 6 months and 5 years depending on contribution record, plus one year of benefit at 70% of minimum wage On expiry of UI, (there is a) possibility of means-tested assistance. The creation of measures of institutions and policies like benefit duration, employment protection, and bargaining coordination requires considerable subjective judgment, and this has raised additional concerns. If the empirical tests are designed to confirm strongly held theoretical priors (institution-caused rigidities explain unemployment) and the same researchers generate the measures of the key explanatory variables (the institutions), it would be likely that measures that do the best job of confirming the guiding hypotheses will be preferred. Blanchard and Wolfers warn of this Darwinian effect: One must worry however that these results are in part the result of research Darwinism. The measures used by Nickell have all been constructed ex-post facto, by researchers who were not unaware of unemployment developments. When constructing a measure of employment protection for Spain, it is hard to forget that unemployment in Spain is very high Also, given the complexity in measuring institutions, measures which do well in explaining unemployment have survived better than those that did not (Blanchard and Wolfers, 2000, p. c22). The 1994 Jobs Study triggered a major OECD effort to produce better quality institutional measures. The objective was, like the first generation efforts, to facilitate tests of the orthodox cornerstone of the Jobs Study, that strong labor market institutions explains employment performance across countries. But the creation of these improved measures took time, and meanwhile researchers demanded longer time series.

11 11 In the case of a measure of the strictness of employment protection laws, Blanchard and Wolfers took the recently developed OECD country estimates (OECD, 1999) and merged them with an entirely different series produced by Lazear (1990) to create an EPL score for each 5-year period from the early 1960s to the mid-1990s. The extraordinarily detailed and carefully constructed OECD EPL measures were available for just two data points: the late 1980s and the late 1990s. From these two estimates for each country, Blanchard and Wolfers created another for by interpolating between the OECD s late 1980s and late 1990s scores, and still another for the early 1980s simply by using the late 1980s figures, on the grounds that they did not have information to suggest that there were any changes between the early and late 1980s. Thus, from two multi-year averages for the late 1980s and late 1990s (OECD), four 5-year averages were created. Blanchard and Wolfers (2000) then the Lazear data to do the same for the first two decades (1960s and 1970s). But the Lazear and OECD measures are quite different, an issue not addressed by either Blanchard-Wolfers or the many subsequent studies that also relied on this EPL series. For the late 1980s and late 1990s, the OECD (1999) took into account three dimensions of employment protection: procedural inconveniences which the employer faces when trying to dismiss; notice and severance pay provisions; and prevailing standards of and penalties for unfair dismissal. Further, their estimates were designed to cover both white and blue collar workers. In contrast, Lazear's index is narrowly confined to just one of the three OECD dimensions, severance pay and notice, and was further limited to the number of months of severance pay or notice a blue collar worker with ten years of service received upon termination without cause (emphasis added, p ). By merging the OECD and Lazear series, Blanchard and Wolfers produced an EPL measure for 5-year periods from the early 1960s to the late 1990s. Nickell et al. (2001, 2003, 2005) then annualized these data by simple interpolation. For consistency and lack of an alternative, this mongrel EPL measure was then used (either in its annual or 5-year format) for the tests published by many of the most influential subsequent studies, including the IMF (2003), Belot and van Ours (2004), Baker et al. (2004 and 2005), and Baccaro and Rei (2005). In contrast, the Bassanini-Duval (2006) employs what must be a far superior annual EPL series that has recently been generated by the OECD, in large part because they limit the analysis to the post 1982 period.

12 12 Again, due to the efforts of the OECD, the quality and comparability of unemployment benefits data for OECD countries improved dramatically after the mid-1990s. The OECD produces an average gross replacement rate (across family types, income levels, and for different durations of unemployment) for every second year since 1961 and this has become the measure of choice for empirical work in this area. This measure allows researchers to capture both the replacement rate and duration in a single measure of benefit generosity. More recently, the OECD has constructed net replacement rates, which take into account unemployment compensation after taxes and various related benefits. These are far more appropriate than the gross replacement rates for measuring the incentives facing workers. Net replacement rates have been constructed for selected dates between the early 1990s and 2003, but the OECD is only now (2006) coming out with a historical time series. These new net benefit figures are particularly attractive because they will measure generosity relative to average wages, not the average production worker wage (which, with the shift to services, is increasingly misleading). The question is whether the new, much improved measures of benefit generosity will perform as well as the average gross replacement rate in regression tests (section 4). There is some reason for doubt. Currently, these new net replacement rates can be compared over time for just 8 of the countries typically included in cross-country tests, and only for If changes in the standard gross replacement rate are good measures of the change in benefit generosity likely to have major effects on labor supply decisions and wage pressure (and therefore on employment and unemployment rates), they should be closely correlated with the new and superior net replacement rates. It turns out that there is little correspondence between the two. Three countries show changes roughly similar in magnitudes, France (NRR: +4 pts; GRR: +2), Austria (NRR: -2; GRR: -1), and the UK (NRR: -1; GRR: -2). But three other countries show changes in opposite directions: the U.S. (NRR: -6; GRR: +2); Japan (NRR: +2; GRR: -2); and Germany (NRR: +1; GRR: -3). The two remaining countries show huge differences in the size of the change: Italy (NRR: +2; GRR: +15) and Finland (NRR: -9: GRR: 0). 10 These are not results that give us much confidence that when the improved net rates become available for the full set of 19 OECD countries, the measured change in benefit generosity will show a close correspondence to

13 13 changes in the measure that many studies have found associated with changes in unemployment. But even a net replacement rate measure that reflects both levels and the duration of benefit relative to the average wage, other critical features of unemployment benefit systems are left unmeasured. There is no cross-country measure of eligibility, but we know that the share of the unemployed who are receiving unemployment-related benefits varies dramatically across countries. For example, younger workers, who in many countries account for a large part of the unemployment problem and who may be most sensitive to work incentives, may be either ineligible for benefits in the first place (e.g., Spain and Italy) or are eligible only for much lower levels of benefit. Closely related, enforcement of eligibility rules is a critical dimension of benefit generosity, but there is no reliable historical series that can be used for cross-country comparisons. Finally, the bargaining power workers gain from trade union led collective action is poorly measured. The most commonly used measure is union density the share of employees who are union members. But union density is not closely correlated with collective bargaining coverage the share of employees whose wages and employment conditions are set through collective bargaining. The most extreme example is France, which had a union density rate of just 10% in 2000, below that of even the U.S. (13%), but a collective bargaining coverage rate of over 90 percent. Countries with less than 40 percent union density and more than 80 percent collective coverage included Austria, Australia, Portugal Spain and The Netherlands. The coverage measure is much harder to produce, which helps explain why it is available for fewer countries and many fewer years than the union density measure. In any case, it is not clear that a perfectly measured union coverage rate would be a particularly good measure of the power of unions to affect market outcomes. This skepticism is suggested by the fairly robust finding that bargaining coordination is associated with lower unemployment (see section 4), which is usually interpreted as indicating that the bargainers have incorporated the effect of wage bargains on employment in their bargaining objectives. Given these considerable inadequacies in measurement on both sides of the relationship for the unemployment rate as well as the key labor market institutions it might be viewed to be surprising if any statistical fit was uncovered. Interestingly, as Section 4 (Table 3) will

14 14 show, as the quality of the labor market institution measures has improved, the strength of the reported statistical association between these measures and the pattern of unemployment has fallen. This may reflect some combination of Blanchard s Darwinian effect in the early literature (the measures that best produce results consistent with particular theoretical expectations are the ones finally used), the use of improved econometric techniques, and the much greater attention to robustness in recent studies. 3. Correlation Evidence 3.1 Casual Associations Where the conventional wisdom is so dominant that there is no competing account to speak of, the standards for evidence are likely to suffer. This may explain the frequent resort to casual association in making the case for the orthodox rigidity explanation. The OECD s Jobs Study offers an example. Although no evidence of a statistically significant relationship between unemployment benefit generosity and unemployment is presented for any particular point in time, Chapter 5 of the Jobs Study states confidently that increases in a more comprehensive measure of unemployment compensation has typically been followed by an increase in unemployment but usually with a considerable lag (p. 44). Support for this conclusion on lagged effects is provided in Chapter 8, and two kinds of evidence are presented. We consider the first one here, in which unemployment increases are explained by earlier increases in unemployment benefits (the replacement rate). According to Chapter 8 of the Jobs Study (OECD 1994, p. 178): In some countries, there have been major reforms in benefit entitlements which give some more specific idea of how long lags may be. In Canada, entitlements rose in 1972 and unemployment rose unusually in 1978 and more strongly around In Finland, entitlements rose in 1972 and unemployment rose sharply (in contrast to its Scandinavian neighbors) through to 1978; in Ireland, changes increasing entitlements occurred over 1971 to 1985, and its rise in unemployment was particularly large (as compared to other European countries) from 1980 to In Norway, major increases in entitlements occurred in 1975 and 1984 (although also before and after these dates), and unemployment rose exceptionally around Entitlements rose in Sweden in 1974 and in Switzerland in 1977, with major rises in unemployment in 1991 in both cases. These experiences suggest lags between rises in entitlements and later sharp rises in unemployment of 5-10 years for Canada, Ireland and Finland but perhaps 10 to 20 years in Norway, Sweden and Switzerland.

15 15 Such breathtaking leaps in association must require extremely strong theoretical priors. As Manning (1998, p. 144) puts it, I think that we would all agree that this is absurd. In fact, one could write a very similar paragraph relating performance in the Eurovision Song Contest to unemployment. To take another example, Heckman (2003, p. 373) suggests that an important part of the German employment performance problem can be traced to what he terms substantial unemployment net benefit replacement rates (79%), because Germans, like all people, respond to these incentives (not to work). More substantial evidence of a causal relationship running from benefits to unemployment for Germany is not offered. Although it goes unmentioned, Heckman s figure also shows that Denmark (80%), the Netherlands (82%), Switzerland (84%) and Sweden (85%) all had higher net replacement rate generosity than Germany (1995). But unemployment rates for these four high generosity countries have consistently been lower than Germany s since Three of the four (Sweden is the exception, but just barely) have shown lower unemployment rates than the U.S. since the late 1990s, despite a much lower U.S. net replacement rate. Similar reliance on casual association can be found in discussion of the labor market effects of employment protection laws. The OECD s Economic Survey of France (OECD 2005b) notes that employment protection is relatively strict in France and, for this reason, calls for a series of reforms. But no evidence or references are offered to establish that EPL strictness actually helps explain French unemployment, or that the recommended reforms would reduce it. Similarly, the OECD s Economic Survey of the Netherlands (OECD 2005c, p. 25) recommends increasing the responsiveness of employment to economic conditions by easing strict EPL or regular contracts, (and) making real wages even more responsive to unemployment by phasing down unemployment benefit replacement rates as unemployment spells lengthen. As in the French country survey, this policy recommendation for the Netherlands (a very low unemployment country) is made without reference to any evidence on the links between either EPL strictness and employment responsiveness or benefit replacement rates and the responsiveness of real wages to unemployment. 3.2 Simple Correlations

16 16 Between these examples of assumed relationships based, presumably, on theoretical common-sense, and the macroeconometric exercises that we review below, lie simple bivariate correlations, most commonly represented by scatter plots. One would expect that if the expected effects of protective labor market institutions on employment performance are as direct and strong as commonly believed, we should observe some evidence of it with simple correlations. Indeed, scatter plots have been frequently employed to show the links between unemployment and various labor market institutions. For instance, it has been argued that extended duration of generous benefits will have particularly strong effects on long-term unemployment. Layard et al. (1994) put particular emphasis on this source of the unemployment crisis: The unconditional payment of benefits for an indefinite period is clearly a major cause of high European unemployment (p. 92, italics in the original). The authors present a plot of a measure of the maximum duration of benefit in years against the long-term share of unemployment for the mid-1980s (1991, Figure 13; 1994, Figure 13) and remark that all the countries where long-term unemployment has escalated have unemployment benefits of some kind that are available for a very long period, rather than running out after 6 months (as in the USA) or 14 months (as in Sweden) (p. 59). This evidence leads to their conclusion that In countries in which benefits are indefinitely available, employment is much less likely to rebound after a major downwards shock (1991, p. 40; 1994, p. 62). As we noted above, Atkinson and Micklewright (1991) have been quite critical of these indefinite duration measures of benefits. Other examples appear in the published literature. Heckman (2003) presents several scatter plots showing a negative relationship between employment rates and the strictness of employment protection laws. In another example, Blanchard (2004) has used scatter plots of cooperation in labor relations in the late 1990s against unemployment to suggest the importance of the quality of labor relations for labor market performance. The OECD s Jobs Study made frequent use of scatter plot and simple correlation evidence. After the passage linking increasing unemployment to increases in benefit generosity quoted above, Chapter 8 proposes to examine correlations more systematically (p. 178). Here they show scatter plots of cycle-to-cycle changes in unemployment rates and the summary measure of benefit entitlements (p. 180). For three periods ( , ,

17 17 and ) the OECD presents scatter plots for the change in unemployment against the 6- year average ( summary ) benefits level as well as against the change in the benefits measure over the previous cycle. This produces 6 correlation tests. They do this both for a full set of 21 countries and for a reduced set of 14 countries, on the grounds that standardized unemployment data were unavailable for 7 of the 21 countries before the 1980s, resulting in a total of 12 tests. They find that In data for 21 countries, none of the individual correlations are statistically significant at the 5 percent level. For the 14-country data, two of the six tests produce the expected positive correlation: using the level of benefit entitlements measure, and using the change in benefits measure. 11 In sum, this is the correlation evidence that supports the OECD s Chapter 5 contention that increases in unemployment tends to follow increases in unemployment compensation (see above). It turns out that as a general rule, simple cross-country correlations between unemployment and the standard measures of the key labor market institutions offer little support for the orthodox account. Using five-year averages for the 1980s and 1990s for 20 OECD countries, Baker et al. (2005) found no statistical association between unemployment and OECD measures of employment protection laws, unemployment benefit replacement rates, the duration of unemployment benefits, union density or union coverage. To further illustrate what the correlation evidence shows, we present some simple scatter plots of unemployment and various measures of unemployment benefit entitlement generosity. We limit these to benefit entitlements both for reasons of space and because our survey of the regression literature (below) indicates that the benefit system is the single labor market institution with the strongest and most robust unemployment-increasing effects. If unemployment benefit entitlement generosity is one of the key institutions at the root of unemployment, the strongest evidence should appear with the use of the OECD s net replacement measure of unemployment benefits. Net benefit is measured as the after-tax value of unemployment assistance and other social assistance, such as housing and child support. The net replacement rate takes this after-tax measure as a share of after-tax household earnings. If workers are calculating the tradeoff between the dole and work, such an after-tax measure is clearly the most appropriate. Figure 2 shows that there is, indeed, a relationship, but it is perverse: in 2002, more generous after-tax benefits (measured as the overall average over 60 months for two earnings levels and three family types) is associated

18 18 with lower unemployment across these 20 countries. As the figure shows, Italy is an outlier (it offers no benefits after the first phase of unemployment), but even without Italy there is a negative relationship. The figure shows eleven countries (in two groups) with unemployment below the US in 2002 but with net replacement rates more than twice as high (60-80% vs 30%). With data from an OECD paper by John Martin (1996), we also found a negative relationship for 1994/5 (not shown). Long duration of benefits is also expected to help explain high unemployment. Figure 3 shows benefit duration plotted against unemployment, with duration measured as the ratio of the net replacement rate in the 60 th month of benefit receipts to the initial phase on the entitlement (effectively the 1 st month). This duration measure can be greater than one because for the generally smaller group still eligible for assistance in the 60 th month, more kinds of social assistance may be available than in the initial phase. The data are shown for 2001, the most recent data available for short and long term net replacement rates. Figure 3 shows that higher levels of benefit duration are associated with lower unemployment. Spain and Italy offered relatively ungenerous long term benefits but have high unemployment; Ireland, Denmark, the UK and Austria had similar or lower unemployment than the U.S., but much more generous long-term unemployment-linked net benefits. As noted above, Layard et al. (1991, 1994) argue that there is a close fit between benefit entitlement duration and long-term unemployment. Figure 4 shows a plot of long-term unemployment against the same duration measure used in Figure 3. The data again fail to show the predicted positive association between benefit generosity and unemployment. Germany and Belgium show high net benefit duration and high long term unemployment, but Ireland, the UK, New Zealand, Denmark and Austria are at least as generous with much lower shares of long-term unemployment. Italy has no long term benefits, but has the highest level of long-term unemployment. A number of the most influential panel data studies that we survey below have focused on the extent to which changes in labor market institutions can account in a substantial way for changes in the pattern of unemployment across countries (e.g., Nickell et al., 2005). Along these lines, changes in benefit duration generosity might be expected to be associated with changes in long-term unemployment. Figure 5 explores this possibility for the period (the longest period the available data permit). The data show no correlation. The

19 19 Netherlands and Norway experienced large declines in both duration and long term unemployment, but Ireland had the largest decline in unemployment at the same time that it had the largest increase in benefit duration. Although Canada had the largest decline in duration of benefits, its share of long-term unemployment showed little change. The benefits measure used in nearly all of the recent time series regression tests is the average gross replacement rate, for which there are now measures from 1961 to The change in this measure has typically been found to be significantly associated with the change in unemployment (see below). Figure 6 shows that the simple correlation between the percentage point change in unemployment and the gross replacement rate has the expected positive sign. Many countries show small changes in the benefits replacement rate (both up and down) and large changes in unemployment (from -7 percentage points for Ireland to +3 points for Japan). In Section 5 we will return to the question of how much weight ought to be assigned to these changes in the gross replacement rate. 4. Macroeconometric Evidence 4.1 The Consensus View: Centrality of Labor Market Institutions As employment performance across much of Europe worsened, economists turned their attention to the links between institutions, rigidities, and unemployment (Bruno and Sachs, 1984?; Blanchard and Summers, 1986; Lindbeck and Snower, 1988). This early research, in turn, spawned a rapidly growing literature aimed at explaining both cross-country differences in unemployment and the evolution of these differences over time with regression tests. The most influential studies share the same broad conclusion: in the final analysis, the evidence offers support for the orthodox theoretical expectation that labor market institutions have played a key role in cross-country unemployment differences. For example: Thus, with six institutional variables plus the change in inflation, we can explain over 90 per cent of the differences in unemployment between countries (Layard, Nickell and Jackman, 1994, p. 82). The broad empirical conclusions suggest that policy variables (labor market institutions) and the institutional mechanisms of wage determination do matter for the level of structural unemployment as well as for the speed of labour market adjustment in the OECD countries (Scarpetta, 1996, p. 45).

20 20 This paper has identified a number of policy settings and institutional features of the labour market which are associated with high structural unemployment (we) assign significant roles to unemployment benefits, collective bargaining structures, active labour market policies and the tax wedge. It requires strong political will and leadership to convince electorates that it is necessary to swallow all of the (deregulation) medicine and that it will take time before this treatment leads to improved labour market performance and falling unemployment. But the success stories show that it can be done! (Elmeskov et al., 1998, pp ) To sum up, reductions in replacement rates, lower tax wedges, liberalized employment protection regulations, and improved active labor market policies remain essential ingredients of a comprehensive labor market strategy geared to reducing Europe s high structural unemployment rate (IMF, 2003, p. 141). Our results indicate (that) broad movements in unemployment across the OECD can be explained by shifts in labour market institutions (Nickell et al., 2005, p. 22). In the next several sections ( ), we survey a number of influential cross-country econometric studies for the purpose of assessing the consistency and robustness of the findings. The case for treating protective labor market institutions as the principal determinants of high unemployment will be stronger the more consistent the findings are across studies and the more care taken to ensure that the published results are robust. Table 3 provides a summary of the implied effects of changes in eight of the most commonly employed institutional measures from eleven panel data regression studies published since While broadly representative, this is by no means a comprehensive list. In some cases studies were not included because it was difficult to make comparable the implied effects. ayed by labor market institutions in the striking decline in UK unemployment in the 1990s. 4.2 Early OECD Studies In Key Lessons for Labor Market Reforms, Elmeskov, Martin and Scarpetta (1998) (hereafter, EMS) aim to distill the main lessons for labour market reforms from the (country) successes and failures revealed by recent OECD research (p. 1). The authors, three senior OECD economists, note that their econometric work is essentially an update and extension of Scarpetta s (1996) earlier work. As Table 3 shows, both Scarpetta and EMS find a significant effect of EPL and unemployment benefit replacement rates, but differ on union density (significant for Scarpetta but not for EMS), the tax wedge (significant for

21 21 EMS but not Scarpetta) and bargaining coordination (same direction, but the implied effect is twice as large in Scarpetta). These rather substantial differences are not addressed in the EMS paper, which is notable since the results of these exercises have been highly influential for the way other researchers and policy makers understand the sources of poor employment performance. Indeed, Elmeskov et al. (1998, p. 2) point to the key role played by Scarpetta s regression results: The OECD work since 1994 has produced a series of additional publications. This work has enabled the Organization to identify a number of country success stories as well as failures in terms of implementing the OECD recommendations and the resulting labour market outcomes. In assessing the needs for reform, the work has relied heavily on the econometric work of Scarpetta (1996). If this influential work is found to be inconsistent in a substantial way with the same author s work a few years later (in EMS, 1998), it would do readers, particularly researchers and policy makers, a considerable service to highlight and explain the differences. EMS compare their findings only to those of Nickell and Layard (1997), noting that while generally similar, the findings for EPL are inconsistent (Nickell and Layard find no significant effect). 4.3 Other Early Cross-Country Tests In many papers and books published in the 1990s, Nickell (in many cases with Layard) reported results from relatively simple cross-country regressions based on the same grouped data ( , ). The results shown in row 3 of Table 3 for Nickell (1997) are representative. Although all the variables except EPL are strongly significant with the expected sign, his conclusion in this paper is cautious: It is clear that the broad-brush analysis that says that European unemployment is high because European labor markets are rigid is too vague and probably misleading. Baker et al. (2005) explored the sensitivity of the main results in Nickell's influential (1997) paper to newer versions of the institutional variables. We replaced six of the eight institutional variables used in the 1997 tests with improved measures that were employed by the same author in more recent work (Nickell et al., 2003; 2005). We also used alternative measures of union coverage (from Blanchard and Wolfers, 2000) and active labor market policies (OECD). 12 With these newer versions of the same institutional measures, the